We propose Seq-LSTM-Conv, a new method for time series forecasting, which combines LSTM and convolutional neural networks. It generates multiple forecasts for input sequences with different length, in a unified framework, without the need to retrain the model for each sequence length. It then aggregates these forecasts to produce the final forecast using static and dynamic strategies, which improves the robustness of the forecasts and facilitates adaptation to the changes in the time series. We also propose an extension of Seq-LSTM-Conv, based on autoencoders, to integrate data from multiple related time series. The method is evaluated for the task of solar power prediction using multi-source data - solar, weather and weather forecasts. The results showed that Seq-LSTM-Conv was the most accurate method; both the static and dynamic versions of Seq-LSTM-Conv outperformed the single LSTM and convolutional neural network and the other methods used for comparison. The dynamic version, Seq-LSTM-Conv-Dyn, was the most accurate model demonstrating the potential of adaptive aggregation based on recent performance. The use of weather data was beneficial in all cases. The results also showed the effectiveness of the proposed autoencoder-based extension of Seq-LSTM-Conv for multi-source data integration.